impact factor
logo
 

Large-vessel vasculitis

 

Colour Doppler ultrasound of temporal arteries for the diagnosis of giant cell arteritis: a multicentre deep learning study


1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14

 

  1. Vascular Medicine Department, Groupe Hospitalier de la Rochelle Ré Aunis, La Rochelle, France. christophe.roncato@ght-atlantique17.fr
  2. Artificial Intelligence Lab, Météo-France, Toulouse, France.
  3. Rheumatology Department, Groupe Hospitalier de la Rochelle Ré Aunis, La Rochelle, France.
  4. Clinical Research Unit, Groupe Hospitalier de la Rochelle Ré Aunis, La Rochelle, France.
  5. Department of Internal Medicine, CHU Nantes, France.
  6. Department of Internal Medicine, CHU Nantes, France.
  7. Department of Internal Medicine, CHU Nantes, France.
  8. Department of Internal Medicine, CHU Nantes, France.
  9. Université Paul Sabatier, Toulouse, France.
  10. Vascular Medicine Department, Groupe Hospitalier de la Rochelle Ré Aunis, La Rochelle, France.
  11. Medicine and Haematology Department, Centre Hospitalier de Rochefort, France.
  12. Rheumatology Department, Groupe Hospitalier de la Rochelle Ré Aunis, La Rochelle, France.
  13. Rheumatology Unit, CHU de Poitiers, France.
  14. Department of Internal Medicine, CHU Nantes, France.

CER13208
2020 Vol.38, N°2 ,Suppl.124
PI 0120, PF 0125
Large-vessel vasculitis

Free to view
(click on article PDF icon to read the article)

PMID: 32441644 [PubMed]

Received: 13/02/2020
Accepted : 16/04/2020
In Press: 21/05/2020
Published: 21/05/2020

Abstract

OBJECTIVES:
Giant cell arteritis (GCA) is the most common systemic vasculitis in adults. In recent years, colour Doppler ultrasound of the temporal arteries (CDU) has proven to be a powerful non-invasive diagnostic tool, but its place in the diagnosis of GCA remains to be defined. A limitation of the CDU is the inter-operator reproducibility. Image analysis from a different perspective is now possible with the development of artificial intelligence algorithms. We propose to assess this technology for the detection of the halo sign on CDU images.
METHODS:
Three public hospitals retrospectively collected data from 137 patients suspected of having GCA between January 2017 and April 2019. CDU images (n=1,311) were labelled with the VIA software. Three sets (training, validation and test) were created and analysed with a semantic segmentation technique using a U-Net convolutional neural network.
RESULTS:
The area under the curve (AUC) was 0.931 and 0.835 on the validation and test set, respectively. An image positivity threshold was determined by focusing on the specificity. With this threshold, a specificity of 95% and a sensitivity of 60% were obtained for the test set. The analysis of the false interpretation showed that the acquisition modalities and the presence of thrombus caused confusion for the algorithm.
CONCLUSIONS:
We propose an automated image analysis tool for GCA diagnosis. The 2018 EULAR guidelines for image acquisition must be respected before generalising this algorithm. After external validation, this tool could be used as an aid for diagnosis, staff training and student education.

Rheumatology Article